Metaheuristic based relevance feedback optimization with support vector machine in content based image retrieval
In this era of information technology, critical fields such as forensic and medical science generates large amount of images. This rapid increase in the digital contents (images) has made Content Based Image Retrieval (CBIR) an attractive research area in the domain of Multimedia. In conventional CB...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Published: |
2015
|
| Subjects: | |
| Online Access: | http://eprints.uthm.edu.my/8043/ http://eprints.uthm.edu.my/8043/1/MUHAMMAD_IMRAN.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In this era of information technology, critical fields such as forensic and medical
science generates large amount of images. This rapid increase in the digital contents
(images) has made Content Based Image Retrieval (CBIR) an attractive research area
in the domain of Multimedia. In conventional CBIR, low level features consisting
of Color, Texture and Shape are used to search relevant images. However, these low
level features are unable to search the similar images as per user semantics which
is known as the gap between low level feature and user semantics Bridging this gap
between the low level features and high level semantics, is one of the most important
challenges for the CBIR. To solve this problem, Relevance Feedback (RF) coupled
with Support Vector Machine (SVM) has been applied. However, when the size
of positive samples marked by the user is small, the performance of CBIR is often
unsatisfactory. To improve the performance of RF for CBIR, this thesis has proposed
a new low level feature extraction technique named as CLD-cw and two new image
retrieving techniques named as PSO-SVM-RF and PSOGA-SVM-RF, which combines
RF and SVM with metaheuristic algorithms called Particle Swarm Optimization (PSO)
and Genetic Algorithm. To prevent PSO from premature convergence, this thesis also
proposed a Laplace mutated PSO. The aim of these new techniques is to minimize
user interaction with the system by minimizing the number of RF. PSO-SVM-RF
and PSOGA-SVM-RF were tested on coral photo gallery containing 10908 images.
Precision, recall and F-Score were used to evaluate the proposed techniques. For
the purpose of validation, the performance of developed approaches was compared
with the performance of other well known CBIR techniques. This comparison was
carried out based on precision and F-score. The experiments showed that PSO-SVMRF
and PSOGA-SVM-RF achieved more than 30% accuracy in terms of precision than
previous CBIR techniques. PSO-SVM-RF and PSOGA-SVM-RF also achieved higher
value of precision and F-Score in less number of RF. |
|---|